Implementation of a Scalable Trace Analysis Tool

Abstract

Monitoring is a way to understand a parallel program. It consists in generating and gathering the information concerning a parallel execution. This method has the drawbacks of being intrusive, greedy for memory and furthermore, the amount of generated data is at least linear in the processor number. With the massively parallel machines the management of the data become impossible, and the traditional and sequential treatment is no more feasible. In this paper, I present a tool for analyzing the large amount of data with a scalable approach. This tool is called PIMSY (for Parallel Implementation of a Monitoring SYstem). It gives as fast as possible the information asked by clients and must be scalable and extensible. To be scalable, PIMSY is composed of a collection of trace servers that deal with a part of the information. To be extensible, each trace server communicates with the others to gather the total information and to give it back to the clients. Our first step does not take care of the generation of trace data, but the future versions of PIMSY will include this feature.

This work was supported by the Ministère de la Recherche et de l’Espace under grant MRE-974

supported by ARCHIPEL SA and CIFRE under contract 920-335